Problem Diagnosis

On November 3, 2026, Advocate General (AG) Maciej Szpunar of the Court of Justice of the European Union (CJEU) issued a non-binding opinion in Like Company v. Google Ireland Limited (Case C-250/25), the first preliminary ruling request before the CJEU to address generative artificial intelligence and copyright law (IPWatchdog; CMS Law). Referred by the Budapest District Court (Budapest Környéki Törvényszék) on April 3, 2025, the dispute examines whether Google infringed the press-publisher related rights of Like Company, a Hungarian web portal operator, between June 13, 2023, and May 31, 2024, through its Bard and Gemini chatbots (IPWatchdog; Verfassungsblog; CMS Law).

For IP managers and legal operations leaders, the opinion exposes a structural flaw in treating generative AI compliance as a single training-data problem. Corporate legal teams have focused primarily on whether static pre-training datasets qualify for statutory exceptions, such as the text and data mining (TDM) exception under Article 4 of the Copyright in the Digital Single Market (CDSM) Directive. However, modern enterprise AI products rely heavily on Retrieval-Augmented Generation (RAG), a process that fetches live web content in real time to ground answers to user prompts (IPWatchdog).

AG Szpunar draws a strict legal boundary between pre-training a language model and executing real-time RAG queries (IPWatchdog). While commercial pre-training can in principle shelter under the Article 4 CDSM TDM exception, real-time retrieval and synthesis directly implicate the exclusive rights of reproduction (Article 2 of the InfoSoc Directive 2001/29/EC), making available to the public (Article 3(2) of the InfoSoc Directive), and the press publishers' related right under Article 15 of the CDSM Directive (IPWatchdog; Verfassungsblog). Because the TDM exception does not cover generative outputs that exceed "individual words or very short extracts," AI service providers cannot rely on user prompts as a liability shield when their systems bypass technical restrictions or output recognizable protected expression (IPWatchdog; Verfassungsblog).

Analysis

Factual Baseline and Procedural Posture

Like Company operates Hungarian news portals funded by advertising revenue and keeps its articles freely accessible online without a paywall (IPWatchdog; CMS Law). In July 2023, one portal published an article regarding Hungarian singer Kozsó and his reported plan to introduce freshwater dolphins to Lake Balaton (IPWatchdog). When a user later prompted Gemini in Hungarian to summarize the article from the publisher's website, the chatbot produced a summary that repeated the core premise while adding fabricated details ("hallucinations") absent from the source text, including proposed dolphin pool dimensions, public reactions, and environmental effects (IPWatchdog; CMS Law).

The Budapest court referred four questions to the CJEU covering both model training and chatbot outputs (Verfassungsblog). Procedurally, AG Szpunar treated the training questions as hypothetical because the evidentiary record before the referring Hungarian court did not establish that Like Company's articles were actually used to train Gemini's underlying weights (IPWatchdog). Nonetheless, he provided a provisional legal analysis on training alongside his primary framework governing RAG (IPWatchdog).

Stage 1: Pre-Training and the Conditional TDM Safe Harbor

In his provisional view, AG Szpunar confirmed that training a large language model (LLM) involves technical reproductions of works, as text must be copied, split into tokens, and converted into mathematical vectors (IPWatchdog; Verfassungsblog). Under CJEU precedent (Infopaq), Article 2 of the InfoSoc Directive harmonizes reproduction broadly to include temporary and partial digital copies, capturing snippets as short as eleven consecutive words if they reflect the author's own intellectual creation (Verfassungsblog).

However, AG Szpunar concluded that the EU TDM exception under Article 4 of the CDSM Directive can, in principle, cover these training reproductions for commercial model development (IPWatchdog). This protection is subject to three strict conditions:

  • Lawful Access: The training material must be lawfully accessible online (IPWatchdog).
  • No Opt-Out Reservation: The rightsholder must not have expressly reserved its TDM rights in an appropriate, machine-readable format under Article 4(3) CDSM (IPWatchdog; Verfassungsblog).
  • Output Memorization Mitigation: Crucially, AG Szpunar added a behavioral condition: developers invoking the TDM exception must take technical steps during development to reduce the risk that their models memorize and generate outputs closely resembling specific training works, and national courts should evaluate whether developers implemented such safeguards when deciding if the exception applies (IPWatchdog).

Stage 2: Real-Time RAG Retrieval and Output Infringement

On the central dispute regarding RAG, AG Szpunar classified real-time retrieval and synthesis by an AI chatbot as an online use of press publications by an information society service provider under Article 15 of the CDSM Directive (IPWatchdog). He proposed a two-part test examining what the chatbot retrieves and what it returns (IPWatchdog):

  • Unlawful Retrieval (Input Stage): Technical copying during RAG retrieval infringes the publisher's right if the chatbot accesses the source article unlawfully by circumventing "effective technical measures" (within the meaning of Article 6(3) of the InfoSoc Directive), such as a paywall or bot-specific anti-scraping blocks (IPWatchdog).
  • Recognizable Expression (Output Stage): A chatbot response constitutes a reproduction under Article 2 of the InfoSoc Directive where the protected elements of the press publication remain "recognisable in that answer, through a side-by-side comparison," enabling users to read the AI summary instead of visiting the original site (IPWatchdog). Furthermore, if answers to related queries systematically reproduce recognizable parts of an article so users can dispense with the original, the chatbot infringes the right of making available to the public, though a single output does not establish systematic access (IPWatchdog).

Conversely, AG Szpunar clarified that press publishers' rights protect economic investment rather than underlying facts (IPWatchdog). If a chatbot extracts bare facts via RAG and presents them in a substantially different form of expression, it does not infringe Article 15 CDSM (IPWatchdog).

Google argued that leaving articles freely accessible on the open web for search indexing implies consent for AI chatbot ingestion, and that any reproduction is triggered by the user's prompt rather than the provider (IPWatchdog; CMS Law). AG Szpunar rejected both arguments (IPWatchdog). First, because publishers rely on open web access for search engine visibility and advertising traffic, failing to deploy technical blocks does not equal implied consent unless quite specific circumstances exist (IPWatchdog). Second, infringing outputs are attributable to the chatbot provider if infringing reuse is built into the service's business model or if the provider fails to adopt diligent operator mitigation measures, such as honoring bot restrictions and filtering closely matching outputs (IPWatchdog). Applying this test to the record, AG Szpunar noted that Like Company had not installed effective anti-scraping blocks, the dolphin summary did not clearly reproduce the article's specific expression, and Google appeared to have mitigation measures in place, leaving final factual verification to the Budapest court (IPWatchdog).

Regulatory Convergence and the Korea Angle

This copyright attribution test mirrors the CJEU's broader jurisprudence on automated systems, including Nacionalinis (C-683/21, assigning joint GDPR controllership to entities commissioning IT tools unless they expressly object before public release) and Dun & Bradstreet Austria (C-203/22, requiring courts and supervisory authorities to balance algorithmic transparency under Article 15(1)(h) GDPR against third-party trade secrets), alongside the EU Product Liability Directive (Directive (EU) 2024/2853, applicable from October 2026, treating AI software as a product subject to strict liability) (Covington / PL&B; European Law Blog).

For Korean technology enterprises and content companies operating under Korea's AI Act (adopted in January 2025 and entering into force in January 2026), AG Szpunar's framework creates asymmetrical outcomes depending on which side of the dispute a company occupies (Covington / PL&B). Korean AI developers deploying LLM search or RAG assistants into the EU market benefit from the confirmation that commercial pre-training falls under Article 4 CDSM and that synthesizing bare facts without original phrasing is lawful. Conversely, Korean media syndicators and digital publishers acting as rightsholders face a higher enforcement burden: as Like Company's evidentiary shortfall illustrates, a rightsholder cannot win an RAG infringement claim on a single factual summary if it failed to deploy active bot-blocking protocols (IPWatchdog).

Strategic Implication

While the CJEU judges are not bound by the Advocate General's opinion and a final judgment will take additional time, corporate IP and legal-ops teams must adjust both offensive content protection and defensive AI deployment practices immediately (IPWatchdog; CMS Law).

  • For AI Developers and Enterprise Deployers (Defensive Compliance): Implement a two-layer technical filter on all EU-facing RAG and search-assistant pipelines to establish "diligent operator" status (IPWatchdog). At the retrieval layer, configure web crawlers to verify and strictly obey both TDM reservation metadata (Article 4(3) CDSM) and bot-specific access blocks (Article 6(3) InfoSoc Directive) without separated fallback scraping. At the generation layer, deploy an automated n-gram or semantic similarity guardrail that compares retrieved press source text against drafted chatbot responses prior to delivery, forcing an abstractive rewrite whenever verbatim overlap exceeds the Article 15 CDSM "very short extract" threshold or preserves recognizable expression (IPWatchdog; Verfassungsblog). Document these anti-memorization and output-filtering controls in internal compliance logs now, as national courts will examine them to determine eligibility for the Article 4 TDM training exception (IPWatchdog).
  • For Content Owners and Publishers (Offensive Portfolio Enforcement): Do not rely on website terms of service or the absence of implied consent to stop AI ingestion. Immediately audit public web portals to ensure machine-readable TDM opt-outs are paired with active, bot-specific technical access restrictions targeting RAG user-agents, while whitelisting standard search-indexing crawlers to preserve ad traffic (IPWatchdog). When preparing infringement litigation in EU national courts, build an evidentiary record using multi-prompt testing across related queries to demonstrate that the target chatbot systematically reproduces recognizable expression to substitute for the original publication, rather than filing suit based on a single summarized output (IPWatchdog).

Sources